What the 0DTE literature actually claims

25 papers, sorted by whether a tick tape with a signed dealer book can test them

Published 2026-08-24 · Sample a literature sweep run 2026-08-02, 25 entries across five topics · Scope empirical papers on US index options and index intraday dynamics · Status a map of other people's claims, not a measurement of ours

Two things make the published record on 0DTE options hard to use. Most of the popular claims about dealer gamma are not in the literature at all — they circulate as charts with no sample attached. And most of the claims that are in the literature cannot be checked by a reader, because checking them needs data that costs money, or because the paper itself is behind a paywall that will not open. This post is the sorting exercise: every paper the sweep found, its central claim in one clause, and a verdict on whether the claim can be run against a tick tape — with the reason, and with the identifier so anyone can go and read the thing.

The sweep was run on 2026-08-02. It has no numbers of its own; every number below is the paper's, reported as the paper reports it.

What "testable" assumes

The verdict column is not "is this true". It is "could this claim be run, end to end, against the following data": a tick SPXW tape with the prevailing NBBO at every print; 5-minute full-chain quotes; per-strike open interest; a Black-Scholes surface carrying gamma, vanna and charm per contract per minute; a signed dealer-inventory book; per-second dealer-gamma levels; and the index recovered by put-call parity. That is what this terminal builds each session — the finished ones are at /sessions — and the tick layer under it is licensed, not free (see Reproduce it). A paper marked not testable is usually not wrong; it is about a different instrument, or it reports no effect size to clear.

1. Conditional variance risk premium

2. Intraday realised-volatility forecasting

This is the thinnest area in the sweep, and the thinness is itself the finding. No published paper was found that forecasts intraday realised volatility at a 30-minute horizon from option state and benchmarks it against trailing RV. The literature splits into next-day forecasting from short-dated implied vol, and high-frequency measurement of spot volatility from options, which is estimation rather than forecasting. An internal result in this shape therefore has no published competitor to beat and no published corroboration. That is a gap, not a validation.

3. 0DTE specifically

The two substantive 0DTE papers — Vilkov, and Almeida/Freire/Hizmeri — sit in section 1, because their central claim is about conditional variance risk premium rather than about 0DTE as such.

4. Dealer inventory and positioning

5. Expiry-day and pinning effects

The count

25 entries. Ten are testable as stated, ten are testable only in part — usually because the instrument is wrong but the claim survives as an analogue, or because the paper reports no effect size to clear — and five are not, because they are cross-sectional equity-option papers, a benchmark rather than a candidate, or a claim about flow mix with nothing to run. Ten carry a [pre-2020] flag, meaning the sample ends before daily SPXW listings began on 2022-05-11. Only Vilkov, Almeida et al., Elms, Albers, Alexiou et al., Chong-Todorov and the gated 0DTE-impact papers use samples that include the daily-expiry regime at all.

What to test first, and why

The sweep ranks its own candidates. Reproduced in order:

1. Muravyev & Ni (2020) — the day/night decomposition. A −1.0%/day overnight versus +0.3%/day intraday split is a large, mechanical, unambiguous claim, priceable at real bid/ask and extendable a decade past their sample. If it holds, it says in one line why intraday short-vol variants keep dying, and that 0DTE — which has no overnight leg — is the worst possible vehicle for the premium. Highest information per unit of work.
2. Vilkov (2026) — conditional out-of-sample 0DTE timing. Same instrument, same vendor, same window, published replication code and reference tables, and a cost model close to ours. A rare chance to check a pipeline against someone else's published numbers before trusting either.
3. Almeida, Freire & Hizmeri (2025) — conditional after-cost Sharpe. Their Low-RV versus High-RV split (after-cost Sharpe 0.404–0.470 against 0.004–0.082) is the largest state-conditioning effect anyone reports for index option selling, and their own claim that the edge died after 2022-05-11 is directly falsifiable on the window where modern data is densest. Decisive either way.
4. Muravyev & Pearson (2020) — real execution costs. Not a strategy, but it targets the parameter that kills results. Measuring where actual SPXW prints sit relative to the prevailing NBBO is a day of work, and a 25–60% cut in the assumed cost would reopen several closed files.
5. Muravyev (2016) — inventory-driven imbalance to option returns. The best mapping onto a signed dealer-inventory book, and it predicts option returns rather than index direction.
6. Elms (2026) plus the Barbon-Buraschi amplitude half. Cheap, and the two agree that expiry and gamma concentration widen rather than compress ranges.
7. Albers (2025) on VIX1D, and the Chong-Todorov / Todorov spot-vol machinery. Feature generators, not strategies.
8. Andersen-Fusari-Todorov negative-tail factor. Strong t-statistics, but 2011–2015 and mid-priced.
9. Chordia-Kurov-Muravyev-Subrahmanyam put order flow. Testable, but it is index direction from option flow.
10. Gârleanu-Pedersen-Poteshman, Chen-Joslin-Ni, Aït-Sahalia et al., Dew-Becker et al. Useful for the shape of a signal; none gives an effect size to clear, and all are pre-2020.

The gated rows, and why you should not trust their numbers

SSRN returns 403 to every automated fetch. Four entries in the map were therefore read from title and abstract metadata plus whatever the citing literature says about them, not from the PDF: Beckmeyer, Branger & Gayda (2023) (SSRN 4404704), Bandi, Fusari & Renò (2023) (SSRN 4503344), Adams, Fontaine & Ornthanalai (2024) (SSRN 4881008) and Vasquez, Amaya, Pearson & Garcia-Ares (2025) (SSRN 5113405). Their claims are stated here as the citing literature states them, and their effect sizes are either absent or second-hand. Nothing in those four rows should be trusted until someone opens the PDF by hand. Two more entries are unverified for the same practical reason from a different publisher: the Albers (2025) PDF and the Gârleanu et al. and Chen et al. PDFs refused automated retrieval, so no magnitudes were extracted from them.

Vilkov's SSRN PDF is gated too, but its full text and code are mirrored on GitHub under an MIT licence, which is why that row is the best-documented one in the map.

Where we have already run one

Honesty about linkage: the sweep does not pair any paper with a published post on this site. Where it points to prior work it points at internal results — a dealer-gamma to next-30-minute realised-volatility regression (β = −0.047, p = 0.0013), a killed net-GEX direction study, a charm signal that flipped sign in 2026 — and none of those has been published here. The amplitude side of Barbon & Buraschi, the attenuation side of Adams et al., and Elms's widening-range result all land on the same sign as that gamma-to-RV regression, which is the strongest convergent evidence in the sweep; the regression itself is not on this site, so no link is offered for it. The published measurements here — the touch surface, level hold rates, the OCC dealer-to-dealer floor, daily book turnover and the pre-registered dealer-state test — are separate work, and inventing a correspondence between one of them and a paper in this map would be the exact error this post exists to avoid.

What it does not say

This page contains no measurement. "Testable" is a judgement about data availability, not about whether a claim is correct, well-identified or likely to replicate; several papers marked testable are expected to be rejected on modern data, and the map says so in the row rather than pretending otherwise. The sweep is one pass by one reader on one day and is certainly incomplete — it applied standing filters that exclude a set of already-dead ideas, so the map is not a neutral census of the field. The claims are summarised in one clause each, which is not enough to argue with; the identifiers are given so you can go and read the argument. Where a paper is paywalled, the summary is second-hand and flagged as such. Nothing on this page is investment advice; see the Terms.

Reproduce it

This post has no numbers of its own, so what there is to reproduce is the search. The sweep used, in this order: the arXiv API, OpenAlex, Semantic Scholar, DuckDuckGo HTML results, and direct PDF pulls from author and repository pages. All four indexes are free and take no key. The route:

# arXiv (full text, free) — e.g. Chong & Todorov
http://export.arxiv.org/api/query?search_query=all:%220DTE%22&max_results=50
https://arxiv.org/pdf/2305.04137

# OpenAlex — metadata, DOI, open-access location, citing works
https://api.openalex.org/works?filter=title.search:0DTE%20options

# Semantic Scholar — abstract verification when the publisher is paywalled
https://api.semanticscholar.org/graph/v1/paper/DOI:10.1016/j.jfineco.2018.12.006?fields=title,abstract,year

# Zenodo — the Elms preprint, open PDF
https://doi.org/10.5281/zenodo.19540116

# GitHub — Vilkov's replication code and annotated paper, MIT licence
https://github.com/vilkovgr/0dte-strategies

What refused automated access. SSRN returns 403 to every automated fetch, with no exception found — that is what gates the four flagged rows. The Wiley publisher PDF for Albers (2025) refused automated retrieval, though its DOI landing page is open. Oxford University Press PDFs blocked the fetch for Gârleanu, Pedersen & Poteshman (2009) and Chen, Joslin & Ni (2019), so both were verified from abstract only. Free author or repository copies did work for Cheng (2019), Andersen-Fusari-Todorov (2017), Barbon & Buraschi (2021), Golez & Jackwerth (2012), Christoffersen et al. (2018) and Almeida et al. (2025).

What running the tests would cost. Every "testable" verdict assumes the tick layer described at the top, and that layer is licensed, not free: the SPXW trade and NBBO tick feed from ThetaData (https://thetadata.net) is the product a reader would have to subscribe to in order to redo any of these on prints rather than on midpoints. Vilkov names ThetaData or Massive as his own vendor, which is why his row is the one directly comparable benchmark. The derived per-session layer — levels, the band, the per-minute series — is free: each finished session is published as https://gex.live/snapshots/YYYY-MM-DD.json, and the dates are listed at /sessions. That is enough for the amplitude and pinning rows (Elms, Golez & Jackwerth, the Barbon-Buraschi amplitude half); it is not enough for anything that needs a fill price.

Part of gex.live research. Measured on the free session archive; every session is free to replay.